Financial voucher processing method based on ultra-lightweight feature compression

Abstract

The invention relates to a financial voucher processing method based on ultra-lightweight feature compression, and belongs to the field of financial artificial intelligence. The method comprises the following steps: collecting financial voucher images and text information, constructing a Mini-Fin data set, and dividing the Mini-Fin data set into a training set and a test set; performing data preprocessing on the image data of the training set by using an HSV brightness enhancement method to obtain financial image data; constructing a financial voucher processing model based on ultra-lightweight feature compression, wherein the model comprises a global lightweight feature extractor, a local lightweight feature extractor, a compressed perceptron, an attention module, a classifier and a text extractor; inputting financial image data into the model for training; optimizing the model through a classification loss function and a text loss function to obtain a trained model; and inputting a to-be-detected financial voucher image in the test set into the trained model to obtain a financial voucher processing result. The efficiency of processing large-scale financial data can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of financial artificial intelligence, and in particular relates to a financial voucher processing method based on ultra-lightweight feature compression. Background Art

[0002] In the modern medical field, with the continuous advancement of the digitalization process, hospitals need to process massive amounts of financial voucher data. These data are not only huge in volume, but also highly diverse, containing information in various formats and structures. Therefore, how to efficiently store, transmit and process these data has become an urgent problem to be solved. In the existing technology, there are widespread problems of data redundancy and excessive resource consumption. In order to meet this challenge, feature compression technology was introduced. Feature compression aims to reduce storage space and increase data processing speed by removing redundancy by retaining important feature information in the data. However, traditional feature compression methods such as PCA often rely on complex mathematical models and calculations, and are difficult to apply in resource-constrained environments.

[0003] Ultra-lightweight feature compression technology provides an effective solution. It can significantly reduce the data dimension without significantly losing data feature information through a simplified compression algorithm. This method is particularly suitable for running on devices with limited computing power, such as embedded systems or mobile devices. In addition, with the help of this technology, financial voucher data can be quickly processed and used for real-time analysis and decision support, greatly improving the efficiency of hospital operations. Therefore, the development of a financial voucher processing method based on ultra-lightweight feature compression is precisely to solve the resource and efficiency problems faced in the current medical financial data processing process, so as to realize a more intelligent and efficient medical financial management system. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a financial voucher processing method based on ultra-lightweight feature compression.

[0005] In order to achieve the above object, the present invention is implemented through the following technical solutions: The present invention provides a financial voucher processing method based on ultra-lightweight feature compression, comprising the following steps: S1. Collect images and text information of financial vouchers, construct the Mini-Fin dataset, and divide it into training set and test set; S2. Use the HSV brightness enhancement method to preprocess the image data of the training set to obtain financial image data; S3. Constructing a financial voucher processing model based on ultra-lightweight feature compression, the model includes a global lightweight feature extractor, a local lightweight feature extractor, a compressed perceptron, an attention module, a classifier, and a text extractor; the financial image data is input into the model to train the model; S4. Optimizing the model by using a classification loss function and a text loss function to obtain a trained model; S5. The financial voucher images to be tested in the test set are input into the trained model to obtain the financial voucher processing results.

[0006] Further, step S1 includes: A label file is created for each financial voucher image to record the text information of the financial voucher image; the collected financial voucher images are divided into four categories, including purchase and sales contracts, receipts, invoices, and payment vouchers; four folders are created for the four categories of financial voucher images to be stored independently, named 1, 2, 3, and 4 in sequence; In each category of financial voucher images, 80% of the data is randomly selected as the training set, and the remaining data is used as the test set.

[0007] Furthermore, in step S3, the global lightweight feature extractor: The global lightweight feature extractor includes a first convolutional layer, a ReLU activation function, an average pooling layer, a width maximum calculation layer, a width average calculation layer, a height average calculation layer, a height maximum calculation layer, a second convolutional layer, a third convolutional layer, a softmax() function, a fourth convolutional layer, a ReLU activation function, and a maximum pooling layer; The financial image data Input into the first convolution layer, ReLU activation function, and average pooling layer to obtain the first convolution feature ; The first convolution feature Input into the maximum width calculation layer, average width calculation layer, average height calculation layer, and maximum height calculation layer to obtain the maximum width vector , width average vector , height average vector , maximum height vector ; Set the maximum width vector Use the concat function and the width average vector in the dimension direction of dimension dim=1 Splice to get the maximum vector of double-layer width , the width average vector Use the concat function with the vector with the largest width in the dimension direction of dim=1 Splice to get the double-layer width average vector , the maximum height vector and the height average vector Add in the channel dimension to get the height vector ; The maximum double-layer width vector Input into the second convolutional layer to get the maximum vector of single layer width , the double-layer width average vector Input into the third convolutional layer to obtain the single-layer width average vector , the maximum single-layer width vector and the single layer width average vector Multiply and input the softmax function to get the width matrix ; The width matrix and the height vector Multiply and input into the fourth convolutional layer, ReLU activation function, and maximum pooling layer to obtain the global features .

[0008] Furthermore, in step S3, the local lightweight feature extractor: The local lightweight feature extractor includes a first convolution layer, a ReLU activation function, a first pooling layer, a second convolution layer, a ReLU activation function, and a second pooling layer; The global features Local features are obtained through local lightweight feature extractor .

[0009] Furthermore, in step S3, the compressed perceptron: The compressed perceptron includes a discrete layer, a Huffman coding layer, a convolutional layer, and a ReLU activation function; The global feature and local features After adding, input into the compressed perceptron to obtain the fusion feature .

[0010] Furthermore, in step S3, the attention module: The attention module includes a first attention module and a second attention module; The fusion feature Get the first attention through the first attention module , the formula is as follows: , in, represents the dot product operation, represents element-by-element multiplication, d represents dimensionality reduction, represents the Sigmoid activation function, α, β, and γ represent the first learnable weighting coefficient, the second learnable weighting coefficient, and the third learnable weighting coefficient, respectively. represents the first convolutional layer, represents the second convolutional layer, represents global average pooling, represents the global maximum pooling, , , is the fusion feature Dimensionally matched learnable query matrix, key matrix, and value matrix, , Respectively represent the fusion features The first matrix and the second matrix after linear transformation, r represents the compression ratio; The fusion feature Get the second attention through the second attention module , the formula is as follows: .

[0011] Furthermore, in step S3, the classifier and text extractor: The classifier includes a compression layer, a first fully connected layer, a second fully connected layer, and a max function; First attention After being processed by the classifier, the classification prediction of financial vouchers is obtained ; The text extractor includes a compression layer, a first fully connected layer, a second fully connected layer, and a string output layer; The first attention and second attention After being processed by the text extractor, the voucher text prediction is obtained Furthermore, in step S4, the classification loss function calculates the loss through a cross entropy loss function, and the text loss function calculates the loss through a connectionist temporal classification CTC loss function.

[0012] The present invention also provides a financial voucher processing system based on ultra-lightweight feature compression, comprising: Data acquisition module: used to collect images and text information of financial vouchers, build Mini-Fin dataset, and divide it into training set and test set; Data preprocessing module: used to preprocess the image data of the training set using the HSV brightness enhancement method to obtain financial image data; Model building module: used to build a financial voucher processing model based on ultra-lightweight feature compression, the model includes a global lightweight feature extractor, a local lightweight feature extractor, a compressed perceptron, an attention module, a classifier, and a text extractor; the financial image data is input into the model to train the model; Model optimization module: used to optimize the model through classification loss function and text loss function to obtain a trained model; Financial voucher processing module: used to input the financial voucher images to be tested in the test set into the trained model to obtain the financial voucher processing results.

[0013] The advantages of the present invention are: The present invention realizes comprehensive feature extraction of financial voucher images by fusing global lightweight feature extractors and local lightweight feature extractors, effectively improving the model's ability to understand financial information. At the same time, the introduction of compressed perceptrons and attention modules can further reduce the complexity and computational cost of the model while maintaining the richness of feature information, making the processing of large-scale financial data more efficient. Through the construction of the Mini-Fin data set and feature compression algorithm, this method realizes the simultaneous classification and text prediction of the medical financial voucher images to be detected, providing an innovative and efficient solution for medical financial voucher processing in the field of financial artificial intelligence, and can provide hospitals with fast and accurate financial information processing services. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0015] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] Example 1 In this embodiment, Figure 1 As shown, the present invention provides a financial voucher processing method based on ultra-lightweight feature compression, and the specific steps include: a) Collect images and texts of various hospital financial vouchers and construct the Mini-Fin dataset based on the usage frequency of each type; Specifically, step a) comprises the following steps: a-1) The Mini-Fin dataset collects a variety of financial vouchers from a hospital, including but not limited to invoices, reimbursement forms, bank statements, and other related financial documents. Ensure the breadth and representativeness of the data to cover common types of financial vouchers. Create a label file for each voucher image to record the text information of the voucher. Divide the collected data into four categories: purchase and sales contracts, receipts, invoices, and payment vouchers. Create four folders for independent storage, named 1, 2, 3, and 4.

[0018] a-2) Randomly select 80% of the data in each class in a-1) as the training set, and use the remaining data as the test set.

[0019] b) Use the HSV brightness enhancement method to preprocess the image data of the training set in the Mini-Fin dataset to obtain financial image data ; c) Financial image data Input into the global lightweight feature extractor to obtain the global features ; Specifically, step c) comprises the following steps: c-1) The global lightweight feature extractor consists of the first convolution layer, ReLU activation function, average pooling layer, width maximum calculation layer, width average calculation layer, height average calculation layer, height maximum calculation layer, second convolution layer, third convolution layer, softmax() function, fourth convolution layer, ReLU activation function, and maximum pooling layer.

[0020] c-2) Among them, the convolution kernel size of the first convolution layer is 3*3, the number of convolution kernels is 1, the stride is 1, and the padding is 'same'. The average pooling layer adopts average pooling, the pooling window size is 1*1, and the stride is 1. The width maximum calculation layer calculates the maximum value of each column vector in the width dimension of the current feature map. The width average calculation layer calculates the average value of each column vector in the width dimension of the current feature map. The height average calculation layer calculates the average value of each row vector in the height dimension of the current feature map. The height maximum calculation layer calculates the maximum value of each row vector in the height dimension of the current feature map. The convolution kernel size of the second convolution layer is 1*2, the number of convolution kernels is 256, the stride is 1, and the padding is 'same'. The convolution kernel size of the third convolution layer is 1*2, the number of convolution kernels is 256, the stride is 1, and the padding is 'same'. The convolution kernel size of the fourth convolution layer is 3*3, the number of convolution kernels is 128, the stride is 1, and the padding is 'same'. The maximum pooling layer uses maximum pooling, the pooling window size is 2*2, and the stride is 1.

[0021] c-3) Image data Input into the first convolution layer, ReLU activation function, and average pooling layer to obtain the first convolution feature , the first convolution feature The dimension is 512×512 and the number of channels is 1.

[0022] c-4) The first convolution feature Input them into the maximum width calculation layer, the average width calculation layer, the average height calculation layer, and the maximum height calculation layer respectively to obtain the maximum width vector , width average vector , height average vector , maximum height vector .

[0023] c-5) Set the maximum width vector Use the concat() function with the width average vector in the dimension direction of dim=1 Splice to get the maximum vector of double layer width . The width average vector Use the concat() function with the vector with the largest width in the dimension direction of dim=1 Splice to get the double-layer width average vector . Set the maximum height vector and the height average vector Add in the channel dimension to get the height vector .

[0024] c-6) Set the maximum double-layer width vector Input into the second convolutional layer to get the maximum vector of single layer width . The double-layer width average vector Input into the third convolutional layer to obtain the single-layer width average vector . Set the maximum single-layer width vector and the single layer width average vector Multiply and input the softmax() function to get the width matrix .

[0025] c-7) Set the width matrix and the height vector Multiply and input into the fourth convolutional layer, ReLU activation function, and maximum pooling layer to obtain the global features .

[0026] d) Global features Input into the local lightweight feature extractor to obtain local features ; Specifically, step d) comprises the following steps: d-1) The local lightweight feature extractor consists of the first convolution layer, ReLU activation function, the first pooling layer, the second convolution layer, ReLU activation function, and the second pooling layer. The convolution kernel size of the first convolution layer is 3*3, the number of convolution kernels is 256, the stride is 1, and the padding is 'same'. The first pooling layer uses maximum pooling, the pooling window size is 3*3, and the stride is 1. The convolution kernel size of the second convolution layer is 3*3, the number of convolution kernels is 512, the stride is 1, and the padding is 'same'. The second pooling layer uses maximum pooling, the pooling window size is 2*2, and the stride is 1.

[0027] d-2) Global features Input to the first convolution layer, ReLU activation function, first pooling layer, second convolution layer, ReLU activation function, second pooling layer to get local features .

[0028] e) Global features and local features Input into the compressed perceptron to obtain fusion features ; Specifically, step e) comprises the following steps: e-1) The compressed perceptron is composed of a discrete layer, a Huffman coding layer, a convolutional layer, and a ReLU activation function.

[0029] The discrete layer uses the quantize_dynamic() function in pytorch to dynamically quantize the features, and the Huffman coding layer is constructed using the heapq module in the Python standard library. The convolutional layer has a convolution kernel size of 3*3, a number of convolution kernels of 512, a stride of 1, and uses 'same' padding.

[0030] e-2) Global features and local features After addition, it is input into the discrete layer, Huffman coding layer, convolution layer, and ReLU activation function to obtain the fusion feature .

[0031] f) Fusion features Input into the first attention module to get the first attention ; Specifically, step f) includes: the fusion feature Get the first attention through the first attention module , the formula is as follows: , in, represents the dot product operation, represents element-by-element multiplication, d represents dimensionality reduction, represents the Sigmoid activation function, α, β, and γ represent the first learnable weighting coefficient, the second learnable weighting coefficient, and the third learnable weighting coefficient, respectively. represents the first convolutional layer, represents the second convolutional layer, represents global average pooling, represents the global maximum pooling, , , is the fusion feature Dimensionally matched learnable query matrix, key matrix, and value matrix, , Respectively represent the fusion features The first matrix and the second matrix after linear transformation, r represents the compression ratio; g) Fusion features Input into the second attention module to get the second attention ; Specifically, step g) includes: the fusion feature Get the second attention through the second attention module , the formula is as follows: .

[0032] h) Focus your attention first Input into the classifier to get the voucher classification prediction ; Specifically, step h) comprises the following steps: h-1) The classifier consists of a compression layer, a first fully connected layer, a second fully connected layer, and a max() function.

[0033] h-2) The compression layer will first pay attention to Compressed into a one-dimensional vector along the channel direction. The number of units in the first fully connected layer is 256. The number of units in the second fully connected layer is 4. The max() function adds 1 to the tensor coordinates corresponding to the maximum value in the output of the second fully connected layer, matches the output with the classification table dictionary, and obtains the voucher classification prediction. .

[0034] i) Focus first and second attention Input into the text extractor to get the voucher text prediction ; Specifically, step i) comprises the following steps: i-1) The text extractor consists of a compression layer, a first fully connected layer, a second fully connected layer, and a string output layer.

[0035] i-2) The compression layer will first pay attention to and second attention After addition, they are compressed into a one-dimensional vector along the channel direction. The number of units in the first fully connected layer is 1024. The number of units in the second fully connected layer is 4096. The string output layer matches each element of the output of the second fully connected layer with the basic dictionary table, and the output is the voucher text prediction. j) Using the Adam optimizer to iterate the parameters of the model in steps c) to i) through the classification loss function and the text loss function to obtain the trained model ; Specifically, step j) includes: the classification loss function calculates the loss through the cross entropy loss function. The text loss function calculates the loss through the connectionist temporal classification CTC loss function.

[0036] k) The voucher image to be processed is preprocessed in step b) and then input into the model trained in step j) to output the classification prediction and text prediction .

[0037] Example 2 In this embodiment, we experimentally compared the performance of the method of the present invention with the existing methods in financial voucher processing. This includes key indicators such as classification accuracy, text extraction accuracy, and processing speed. We compared three methods: method A (CRNN model), method B (SwinTR model), and the method of the present invention. The CRNN model implements end-to-end variable-length text recognition and includes three parts: convolutional layer (CNN), recurrent layer (RNN), and transcription layer (CTC loss). SwinTR (Swin Transformer for Text Recognition) is a text recognition model based on Swin Transformer.

[0038] Table 1 Comparison of experimental results As can be seen from Table 1, the classification accuracy of the method of the present invention is significantly better than that of method A and method B, reaching 97.6%, thanks to the effective combination of the optimized feature extractor and the attention module. The text extraction accuracy of the method of the present invention is 95.8%, which is significantly higher than 78.3% of method A and 89.5% of method B, indicating that it has stronger text recognition ability under complex backgrounds. In terms of processing speed, the method of the present invention only takes 300ms to process an image, which is much faster than 1200ms of method A and 3850ms of method B, which reflects the speed advantage of ultra-lightweight feature compression technology.

[0039] In practical applications, the financial voucher processing method of the present invention can be used to automate the classification and text extraction of financial vouchers, thereby improving the efficiency of hospital financial management. The initial data are real-world image samples with different resolutions, brightness, and background noise. The low-resolution blurred instance images in the test set have high performance requirements for the model: after model training, the financial voucher images in the test set are input into the trained model for prediction, and the classification prediction and text extraction results corresponding to the financial vouchers can be obtained.

[0040] Through the above processing methods, it can be clearly seen that the method of the present invention is effective in financial voucher processing, especially in improving the accuracy of text information extraction and reducing the workload of manual review at low resolution; the method of the present invention surpasses other technical solutions in terms of classification and text extraction accuracy and processing speed. This not only improves the automation level of voucher processing, but also reduces the consumption of computing resources, and is very suitable for large-scale medical financial document processing tasks.

[0041] Example 3 This embodiment provides a financial voucher processing system based on ultra-lightweight feature compression, including: Data acquisition module: used to collect images and text information of financial vouchers, build Mini-Fin dataset, and divide it into training set and test set; Data preprocessing module: used to preprocess the image data of the training set using the HSV brightness enhancement method to obtain financial image data; Model building module: used to build a financial voucher processing model based on ultra-lightweight feature compression, the model includes a global lightweight feature extractor, a local lightweight feature extractor, a compressed perceptron, an attention module, a classifier, and a text extractor; the financial image data is input into the model to train the model; Model optimization module: used to optimize the model through classification loss function and text loss function to obtain a trained model; Financial voucher processing module: used to input the financial voucher images to be tested in the test set into the trained model to obtain the financial voucher processing results.

[0042] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A financial voucher processing method based on ultra-lightweight feature compression, characterized in that: The following steps are involved: S1. Collect images and text information of financial vouchers, construct the Mini-Fin dataset, and divide it into training set and test set; S2. Use the HSV brightness enhancement method to preprocess the image data of the training set to obtain financial image data; S3. Constructing a financial voucher processing model based on ultra-lightweight feature compression, the model includes a global lightweight feature extractor, a local lightweight feature extractor, a compressed perceptron, an attention module, a classifier, and a text extractor; the financial image data is input into the model to train the model; S4. Optimizing the model by using a classification loss function and a text loss function to obtain a trained model; S5. The financial voucher images to be tested in the test set are input into the trained model to obtain the financial voucher processing results.

2. The method for processing financial vouchers based on ultra-lightweight feature compression according to claim 1, characterized in that: Step S1 includes: A label file is created for each financial voucher image to record the text information of the financial voucher image; the collected financial voucher images are divided into four categories, including purchase and sales contracts, receipts, invoices, and payment vouchers; four folders are created for the four categories of financial voucher images to be stored independently, named 1, 2, 3, and 4 in sequence; In each category of financial voucher images, 80% of the data is randomly selected as the training set, and the remaining data is used as the test set.

3. The financial voucher processing method based on ultra-lightweight feature compression according to claim 1 is characterized in that: Global lightweight feature extractor in step S3: The global lightweight feature extractor includes a first convolutional layer, a ReLU activation function, an average pooling layer, a width maximum calculation layer, a width average calculation layer, a height average calculation layer, a height maximum calculation layer, a second convolutional layer, a third convolutional layer, a softmax() function, a fourth convolutional layer, a ReLU activation function, and a maximum pooling layer; The financial image data Input into the first convolution layer, ReLU activation function, and average pooling layer to obtain the first convolution feature ; The first convolution feature Input into the maximum width calculation layer, average width calculation layer, average height calculation layer, and maximum height calculation layer to obtain the maximum width vector , width average vector , height average vector , maximum height vector ; Set the maximum width vector Use the concat function and the width average vector in the dimension direction of dimension dim=1 Splice to get the maximum vector of double layer width , the width average vector Use the concat function with the vector with the largest width in the dimension direction of dim=1 Splice to get the double-layer width average vector , the maximum height vector and the height average vector Add in the channel dimension to get the height vector ; The maximum double-layer width vector Input into the second convolutional layer to get the maximum vector of single layer width , the double-layer width average vector Input into the third convolutional layer to obtain the single-layer width average vector , the maximum single-layer width vector and the single layer width average vector Multiply and input the softmax function to get the width matrix ; The width matrix and the height vector Multiply and input into the fourth convolutional layer, ReLU activation function, and maximum pooling layer to obtain the global features .

4. The method for processing financial vouchers based on ultra-lightweight feature compression according to claim 3 is characterized in that: Local lightweight feature extractor in step S3: The local lightweight feature extractor includes a first convolution layer, a ReLU activation function, a first pooling layer, a second convolution layer, a ReLU activation function, and a second pooling layer; The global features Local features are obtained through local lightweight feature extractor .

5. The method for processing financial vouchers based on ultra-lightweight feature compression according to claim 4 is characterized in that: Compressed perceptron in step S3: The compressed perceptron includes a discrete layer, a Huffman coding layer, a convolutional layer, and a ReLU activation function; The global feature and local features After adding, input into the compressed perceptron to obtain the fusion feature .

6. The method for processing financial vouchers based on ultra-lightweight feature compression according to claim 5, characterized in that: Attention module in step S3: The attention module includes a first attention module and a second attention module; The fusion feature Get the first attention through the first attention module , the formula is as follows: ， in, represents the dot product operation, represents element-by-element multiplication, d represents dimensionality reduction, represents the Sigmoid activation function, α, β, and γ represent the first learnable weighting coefficient, the second learnable weighting coefficient, and the third learnable weighting coefficient, respectively. represents the first convolutional layer, represents the second convolutional layer, represents global average pooling, represents the global maximum pooling, , , is the fusion feature Dimensionally matched learnable query matrix, key matrix, and value matrix, , Respectively represent the fusion features The first matrix and the second matrix after linear transformation, r represents the compression ratio; The fusion feature Get the second attention through the second attention module , the formula is as follows: 。 7. The method for processing financial vouchers based on ultra-lightweight feature compression according to claim 6, characterized in that: Classifier and text extractor in step S3: The classifier includes a compression layer, a first fully connected layer, a second fully connected layer, and a max function; First attention After being processed by the classifier, the classification prediction of financial vouchers is obtained ; The text extractor includes a compression layer, a first fully connected layer, a second fully connected layer, and a string output layer; The first attention and second attention After being processed by the text extractor, the voucher text prediction is obtained .

8. The method for processing financial vouchers based on ultra-lightweight feature compression according to claim 1, characterized in that: In step S4, the classification loss function calculates the loss through the cross entropy loss function, and the text loss function calculates the loss through the connectionist temporal classification CTC loss function.

9. A financial voucher processing system based on ultra-lightweight feature compression, characterized in that: include: Data acquisition module: used to collect images and text information of financial vouchers, build Mini-Fin dataset, and divide it into training set and test set; Data preprocessing module: used to preprocess the image data of the training set using the HSV brightness enhancement method to obtain financial image data; Model building module: used to build a financial voucher processing model based on ultra-lightweight feature compression, the model includes a global lightweight feature extractor, a local lightweight feature extractor, a compressed perceptron, an attention module, a classifier, and a text extractor; the financial image data is input into the model to train the model; Model optimization module: used to optimize the model through classification loss function and text loss function to obtain a trained model; Financial voucher processing module: used to input the financial voucher images to be tested in the test set into the trained model to obtain the financial voucher processing results.

Images